QC sampling ratio changed last week; assistant still quotes old numbers. Staff trust AI more—and act on old rules. Governance failure: no version on knowledge, no retirement. When policy changes, slices must replace or go offline; assistant citations need effective dates. Models do not read OA bulletins unless bulletins are a controlled knowledge stream.
Updating knowledge and updating policy are one process with two steps. Skip step two and digitization spreads error faster.
Knowledge Needs Effective and Void Dates
Each policy, SOP, price list entering corpus carries version, effective date, void date, owner. New release makes old version non-retrievable for assistant—audit archive read-only if needed. Hard-coded prompt rules belong on change tickets—no hiding in personal system prompts.
- OA policy publish triggers knowledge task—incomplete means no "everyone trained" claim.
- Answers must link to version—no link, mark unreliable.
- Spot-check with new rules—still hitting old text is an incident.
Retirement Matters as Much as Publish
Adding slices without deleting lets retrieval randomly hit expired content. The XYN digital intelligence system binds policy versions to assistant knowledge—void means invisible to Q&A. Success: policy change day, assistant answers new—not bigger model parameters.
Put knowledge update in the last box of the policy workflow. Without it, intelligence systematically uses old intelligence.
